YC Partner Explains How to Build Self-Improving AI-Native Companies

I was happen to be trying Gbrain from @garrytan for the first time today. This video just came out at the perfect time and helped me understand how Gbrain and the AI native company concept work.

This somehow feels like that early 2000s moment where open source db became a thing and some kids in Harvard decided to build a social network out of it.

Now I finally switched to Obsidian for taking notes, with a vault folder inside /.gstack so Gbrain can see it. Gonna go ahead and integrate email, meetings and more into it!

We’ve been building this at JintellarCore.

A self-improving AI orchestration system that continuously optimizes itself through feedback, routing, memory, and LoRA training loops.

Opening a small beta for teams interested in design partnerships. #AgenticAI #AutonomousAI #EnterpriseAI #AIGovernance #SecureAI

@t_blom @garrytan So interesting. Let’s try this out with http://www.honeyruns.com for fleet management @collin_michaelm

@t_blom I'm curious. How do we solve the compounding effects of hallucination here?

two questions on the recursive loop at scale. how do you keep retrieval landing on correct data versus just plausible data as the brain grows. and how does old data get deprioritized or pruned. a 5-person company and a 50-person one have different "true" facts, and stale data actively misleads. is forgetting automatic or owned by someone?

@t_blom @garrytan I love this and inspired me to start making my own personal automations more public! The future is awesome

https://github.com/DiogoNeves/glide-obsidian

@t_blom Good stuff Tom. We’re leveraging prediction markets to approach this problem space for agents. Some recent achievements

@t_blom Firing agents then the successors learning from what they messed up has been huge

@t_blom It’ll be cool if you share say 10-20 use cases of self improving processes beyond office hours. This could inspire more ideas for founders

@t_blom my bet is in-context learning + a lot of compute, on the order of magnitude of 100+tril tokens to see meaningful improvements

@t_blom This question makes me think of a master craftsman who gets better year over year from deliberate practice, attention, patience, feedback from others, and a dedication to excellence in his craft. That’s a self-improving system.

@t_blom Love the way how you explain it

@t_blom the compounding effects of this are so severely underestimated. it's just hard to appreciate that exponential early on

which giants do you consider will get disrupted the soonest?

@t_blom Love the talk, was very exciting how YC internally has adopted this too

@t_blom @LilysAI_

@t_blom We're building Ren to power every company in the world become L5 autonomous.

@t_blom @garrytan Good one, just watched it - gave me some points to work on more. Great concept.

I’ve experimented recently in this direction and cut 10x+client communication for a new project and honestly I haven’t build still an internal system in some way, just fed my info into a Claude chat

@t_blom easier to self-improve when sam just handed every yc startup $2m in openai tokens. free fuel fr

@t_blom We nailed that sensor layer, Tom.

Using the feedback daily. Added manual recursive loops already.

Results:

6 skills built that reduce tool calls by 70% verifiably.

@t_blom Love the idea. Building a knowledge graph for the company seems like a foundational step in become self-learning and improving. It’ll be interesting to see how complex organizations adopt this principle to evolve.